AMD Phenom II X4 940 & 920: A True Return to Competition
by Anand Lal Shimpi on January 8, 2009 12:00 AM EST- Posted in
- CPUs
45nm and Low Power Consumption
Built on a 45nm process, Phenom II is finally using the same feature size as Intel’s processors. Intel’s 45nm caches are a bit smaller than AMD’s, but it’s no longer a 65nm vs. 45nm playing field - things are much more even. However, AMD and Intel’s approaches to 45nm differ considerably.
Circuits are placed on silicon wafers through the use of photolithography. Light is shone on a mask, and the light then makes it through the mask and etches the circuits on the silicon wafer. The wavelength of light used determines the minimum feature size of the circuits on the wafer. By itself, 193nm wavelength deep ultra-violet light is only useful for circuit feature sizes down to 50nm. To reach 45nm and beyond you need to do a little more.
AMD uses immersion lithography, which places a liquid between the source of the light and the wafer itself. The liquid increases the resolution at which the light can focus, allowing for smaller than 50nm feature sizes with currently available tools. Immersion lithography isn’t a performance enhancing feature; it’s simply one that makes it possible for AMD to manufacture at 45nm.
Intel claims that immersion lithography isn't necessary at 45nm and doesn't use it. Intel uses a technique known as double patterning but only on the gate layer of the chip. Intel’s approach requires higher mask costs but can result in a high yield 45nm chip without the use of immersion lithography. AMD’s approach should be more cost effective initially since you have to create fewer masks, but Intel’s scale of 45nm production should help offset that. For what it's worth, the double patterning has been in use since Intel's 65nm process.
Remember Intel’s high-k + metal gate transistor announcement? That’s still a feature advantage that Intel holds at 45nm. The new transistors make sure that current doesn’t flow when it’s not supposed to, reducing power consumption.
The two processes, despite both being 45nm, are different enough that they aren't the same despite having similar feature size - but comparing manufacturing processes is beyond the scope of this article.
A Power Efficient Phenom?
When Phenom first hit, not only was it underperforming, but it also drew far too much power. Combine that with a CnQ mode that robbed users of performance and you ended up with a CPU that was hardly power efficient. Just like the cache deficiency, Phenom II fixes this.
With Core i7, Intel developed power gate transistors that can completely shut off an individual core that’s not in use. Intel’s cache hierarchy is inclusive so any data stored within a core’s L1 and L2 caches is already duplicated in the L3 cache; if a core isn’t in use it can be shut down and there’s no need to wake it back up until it’s needed again.
Remember, Phenom II isn’t a complete redesign, so AMD couldn’t work on a similar technology. Despite that, idle power in Phenom II is greatly improved. When a single core is idle, the contents of its L1 and L2 can be flushed out to L3, allowing the processor to halt the clocks to that core - thus reducing power. The core will still consume leakage power, but it’ll be far less than if it were running at the lowest p-state. Intel introduced something similar back in the Conroe days, except data from L1 was pushed out to L2 before the core was powered down since there was no L3. Nehalem still has the ultimate in idle power thanks to Intel’s power gate transistors, but as you can see below Phenom II’s idle numbers are quite impressive.
| Processor | Idle Power | Load Power |
| AMD Phenom II X4 940 (3.0GHz) | 109.6W | 189.7W |
| AMD Phenom 9950 BE (2.6GHz) | 124.2W | 210W |
| AMD Phenom X3 8750 (2.4GHz) | 127.5W | 210W |
| AMD Athlon X2 6400 (3.2GHz) | 101W | 195W |
| Intel Core i7-965 (3.2GHz) | 99W | 199W |
| Intel Core i7-920 (2.66GHz) | 95W | 168W |
| Intel Core 2 Extreme QX9770 (3.2GHz) | 135W | 219W |
| Intel Core 2 Quad Q9400 (2.66GHz) | 126W | 174W |
| Intel Core 2 Quad Q8200 (2.33GHz) | 134W | 168W |
| Intel Core 2 Duo E8600 (3.33GHz) | 124W | 157W |
Note that while Penryn’s idle power isn’t nearly as low as Phenom II, this has more to do with Penryn’s lowest operating frequency. Phenom II’s minimum p-state is only 800MHz, compared to 2.0GHz on Penryn. Load power is also impressive, not quite Nehalem impressive but definitely competitive with Penryn at least.
In load power, Penryn still has the advantage. The Q9400 draws 174W compared to 190W for the Phenom II X4 940. The Core i7 is still the most power efficient of the two, as the i7-920 draws less power and is faster than the Phenom II X4 940.
Facebook
Twitter
RSS
93 Comments
View All Comments
Proteusza - Thursday, January 8, 2009 - link
No, I said I hoped it could at least compete with a Core 2 Duo.if its too much to hope that a 2 year younger, 758 million transistor CPU could compete clock for clock with a first gen Core 2 Duo, then AMD has truly fallen to new lows. It has more transistors than i7, and yet it cant compete with a Core 2 Duo let alone i7. What happened to the sheer brilliance of the A64 days? It could beat the pants off any Pentium 4. Now the best AMD can do is barely acceptable performance at a higher clockspeed than Intel needs, all the while using a larger die than Intels.
This keeps them in the game, but it means I wont bother buying one. Why should I?
coldpower27 - Thursday, January 8, 2009 - link
Those days are over, their success was also contigent with Intel stumbling a bit and they did that with P4, with Intel firing on all cylinders, AMD at acceptable is just where they are supposed to be.Denithor - Thursday, January 8, 2009 - link
It wasn't so much of a stumble, more like a face-plant into a cactus. Wearing shorts and a tshirt.Intel fell flat with Netburst and refused to give up on it for far too long (Willamette -> Northwood -> Prescott -> Cedar Mill). I mean, the early days of P4 were horrible - it was outperformed by lower-clocked P3 chips until the increased clockspeed was finally too high for architectural differences to negate.
Into this mix AMD tossed a grenade, the A64 - followed by the X2 on the same architecture. With its IMC and superior architecture there was no way Netburst could compete. Unfortunately, AMD hasn't really done anything since then to follow through. And even today's PII isn't going to change things dramatically for them, they're still playing second fiddle to Intel's products (which means they're forced into following Intel's lead in the pricing game).
JKflipflop98 - Thursday, January 8, 2009 - link
Damn it feels good to be a gangsta ;)Kob - Thursday, January 8, 2009 - link
Thanks for the meaningful comparison with such a wide range of processors. However, I wonder why the benchmarks are so much tilted toward the graphics/gaming world. I think that many in the SOHO world will benefit from test results of other common applications/fields such as VS Compilation, AutoCAD manipulation, Encryption, simple database indexing and even a Chess game.ThePooBurner - Thursday, January 8, 2009 - link
In the article you compare this to the 4800 series of GPUs. I actually see this as the 3800 series. It works out perfectly. The 2900 came along way late and didn't deliver, used to much power, didn't overclock well, and was just all around a looser of a card. Then the 3800 came along. Basically the same thing, but with a die shrink that allowed it to outstretch, just enough, it's predecessor. It was the first card where they got the mix right. After that came the 4800 with a big boost and even more competition. This is what i now see happening with the CPU line. The Phenom 1 was the 2900, and the Phenom II is the 3800. Getting the mix right and getting ready for the next big swing. But, as you point out, Intel isn't likely to sit back, and we can all agree that they are a much different competitor than Nvidia is.Denithor - Thursday, January 8, 2009 - link
...and just like the 3800 series, it falls just short of the target.Remember? The 3870 couldn't quite catch the 8800GT and the 3850 couldn't quite match the 9600GT. While they weren't bad cards, they unfortunately also didn't give AMD the muscle to set pricing where they wanted it, instead they had to put pricing in line with how nVidia priced their offerings.
Same is happening here, with AMD pricing their chips in line with Intel's Q9400/Q9300 processors. And they may have to drop those prices if Intel cuts the Q9550/Q9400 down another peg.
Griswold - Friday, January 9, 2009 - link
Rubbish theory. First of all, these cards were actually available whereas the 8800GT was in extreme short supply and thus much more expensive for many weeks, even into 2008, because it literally made everything else nvidia had to offer obsolete. I couldnt get one and settled for a 3870 for that reason.Secondly, the 9600GT? Do you realize how much later that card came to the game than the 3850? It hit the market near the end of february. Thats almost 3 months after the launch of the 38xx part.
The whole comparison is silly.
ThePooBurner - Friday, January 9, 2009 - link
The 3800 line wasn't ever meant to beat the 8800 line. It just wasn't in the cards. It's purpose was to get the reins back under control. Cut the power and get back to a decent power/performance ratio as well as get equal power to a previous generation in a smaller package to help improve margins. It was a stage setter. From the first time i read about it i knew that it was just a setup for something more, something "bigger and better" that was going to come next. And then the 4800 came along and delivered the goods. I get this same feeling reading about the Phenom II. It's setting the stage. Getting about the same power (a small bump, just like the 3870 over the 2900) in a smaller package, a better power/performance ratio, etc.. This is simply a stage setting for the next big thing. The next CPU from AMD after this one is going to deliver. I'm sure of it.Kougar - Thursday, January 8, 2009 - link
If you tried Everest and Sandra, what about CPU-Z's cache latency tool? It's not part of the CPU-Z package anymore, but they still offer it. Link: http://www.cpuid.com/download/latency.zip">http://www.cpuid.com/download/latency.zipI thought this tool was very accurate, or is this not the case? It even detected the disabled L3 cache on a Northwood that turned out to be a rebadeged Gallatin CPU.